AI & Agents

What’s New in OpenCode 2.3.2: A GPT-6.2 Prompt-Caching Fix

Leave a comment

OpenCode 2.3.2 is a small 2.x point release after 2.3.0, and it is aimed at a billing bug that showed up once GPT-6.2 arrived. The release notes say GPT-6.2 models do not cache a system prompt sent in the Responses API instructions field, so every step was re-billed as uncached input. OpenCode now sends that prompt as a leading developer message on every Responses route, including OpenAI (API key and ChatGPT sign-in), GitHub Copilot, Zen, Azure, and OpenRouter, so it can be cached from the second step on. On GitHub Copilot’s gpt-6.2-sol, OpenCode reports that uncached input per step dropped from about 14,000 tokens to 3. Compaction through /responses/compact now reads the same cached prefix. A smaller fix: webfetch structured output no longer repeats the url and format it was called with.

This is not the GPT-6.2 model announcement itself. It is the CLI learning how that model’s cache actually behaves. I have not upgraded for this post, and I have not measured the token drop myself.

Review before you install: read the release notes yourself, back up your config, and try new versions somewhere you can afford to break first.

My take

If you already pointed OpenCode at GPT-6.2 and your input token counts looked stuck on the uncached side, this is the release to read. A three-token uncached prefix versus 14,000 is a vendor-reported example, but the underlying API mismatch is the kind of thing you want fixed quickly. Read the notes, then bump a copy you can afford to break.

God Bless, and May the Source Be with You!

References & Further Reading

Subscribe (RSS) Follow AI & Agents via RSS

Leave a note

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.